HED-Net: a hybrid ensemble deep learning framework for breast ultrasound image classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer diagnosis relies heavily on medical imaging, with ultrasound being a key modality.
• Accurate and efficient interpretation of breast ultrasound images is crucial for timely diagnosis and treatment.
• Deep learning offers potential for automating and improving the accuracy of image analysis.

Purpose of the Study:

• To develop and evaluate a hybrid deep learning ensemble framework for breast ultrasound image classification.
• To improve the accuracy and efficiency of breast cancer diagnosis using artificial intelligence.
• To enhance the interpretability and robustness of deep learning models in medical imaging.

Main Methods:

• A hybrid deep learning ensemble framework, HED-Net, was developed, combining EfficientNetB7, DenseNet121, and ConvNeXtTiny convolutional neural network models.
• Individual models were trained in parallel to extract diverse features (local, structural, global).
• Feature fusion was achieved using XGBoost, with soft voting ensemble for probability averaging, and SHAP/Grad-CAM for interpretability.

Main Results:

• The HED-Net framework achieved high performance across multiple datasets (BUSI, BUS-UCLM, UDIAT).
• For instance, on the UDIAT dataset, accuracy reached 96.97%, precision 100.00%, recall 90.91%, F1 score 95.24%, and AUC 99.17%.
• Model interpretability was enhanced through SHAP and Grad-CAM visualizations.

Conclusions:

• The proposed HED-Net framework demonstrates significant potential for clinical application in breast ultrasound image classification.
• It offers enhanced diagnostic accuracy and reduced interpretation time, particularly valuable in resource-limited settings.
• The model's robustness and transparency are improved through advanced visualization techniques.